Work

Utility-Aware Retention Policies for Bounded Telemetry Buffers: Trade-offs in Anomaly-Preserving Sampling on the SKAB Benchmark

Year: 2026
Type: preprint
Abstract: Telemetry systems with fixed storage must decide which samples to keep. This work studies utility-aware retention policies that score each arriving sample by recency, anomaly strength, deviation from a running baseline, and rarity, and keep the highest-utility samples in a bounded buffer. We compare ten policies (FIFO, LIFO, random eviction, reservoir, time-stratified, weighted reservoir, value-biased, utility-aware, stratified utility, and a tuned variant) on the SKAB industrial benchmark under a strict chronological protocol: policies see only the first 60% of the stream, weights are tuned on a validation segment, and detectors trained on the retained samples are evaluated on a held-out later period over 10 seeds and four buffer sizes. Utility-aware retention improves unsupervised Isolation Forest performance, especially at small buffers, and matches uniform sampling for a supervised classifier, but it covers fewer anomaly events and distorts the data distribution relative to reservoir sampling. We report these trade-offs, ablations, significance tests, and per-arrival latency and memory costs.
Language: English
Open Access status: green
OpenAlex ID: W7225991730
Cites:
Cited by:
Related to: 10